xref: /freebsd/sys/kern/sched_ule.c (revision e7dd6e9402cae324c2190a70081854c3c8a8feb9)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * This file implements the ULE scheduler.  ULE supports independent CPU
31  * run queues and fine grain locking.  It has superior interactive
32  * performance under load even on uni-processor systems.
33  *
34  * etymology:
35  *   ULE is the last three letters in schedule.  It owes its name to a
36  * generic user created for a scheduling system by Paul Mikesell at
37  * Isilon Systems and a general lack of creativity on the part of the author.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
64 #include <sys/umtx.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
67 #include <sys/sbuf.h>
68 
69 #ifdef HWPMC_HOOKS
70 #include <sys/pmckern.h>
71 #endif
72 
73 #ifdef KDTRACE_HOOKS
74 #include <sys/dtrace_bsd.h>
75 int				dtrace_vtime_active;
76 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
77 #endif
78 
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
81 
82 #define	KTR_ULE	0
83 
84 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 
88 /*
89  * Thread scheduler specific section.  All fields are protected
90  * by the thread lock.
91  */
92 struct td_sched {
93 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
94 	short		ts_flags;	/* TSF_* flags. */
95 	int		ts_cpu;		/* CPU that we have affinity for. */
96 	int		ts_rltick;	/* Real last tick, for affinity. */
97 	int		ts_slice;	/* Ticks of slice remaining. */
98 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
99 	u_int		ts_runtime;	/* Number of ticks we were running */
100 	int		ts_ltick;	/* Last tick that we were running on */
101 	int		ts_ftick;	/* First tick that we were running on */
102 	int		ts_ticks;	/* Tick count */
103 #ifdef KTR
104 	char		ts_name[TS_NAME_LEN];
105 #endif
106 };
107 /* flags kept in ts_flags */
108 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
109 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116     sizeof(struct thread0_storage),
117     "increase struct thread0_storage.t0st_sched size");
118 
119 /*
120  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129 
130 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
131 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
134 
135 /*
136  * Cpu percentage computation macros and defines.
137  *
138  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
139  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
140  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
141  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
142  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
143  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
144  */
145 #define	SCHED_TICK_SECS		10
146 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
147 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
148 #define	SCHED_TICK_SHIFT	10
149 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 
152 /*
153  * These macros determine priorities for non-interactive threads.  They are
154  * assigned a priority based on their recent cpu utilization as expressed
155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
156  * and end of the MIN to MAX timeshare range are only reachable with negative
157  * or positive nice respectively.
158  *
159  * PRI_RANGE:	Priority range for utilization dependent priorities.
160  * PRI_NRESV:	Number of nice values.
161  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
162  * PRI_NICE:	Determines the part of the priority inherited from nice.
163  */
164 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
165 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
166 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define	SCHED_PRI_TICKS(ts)						\
170     (SCHED_TICK_HZ((ts)) /						\
171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define	SCHED_PRI_NICE(nice)	(nice)
173 
174 /*
175  * These determine the interactivity of a process.  Interactivity differs from
176  * cpu utilization in that it expresses the voluntary time slept vs time ran
177  * while cpu utilization includes all time not running.  This more accurately
178  * models the intent of the thread.
179  *
180  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
181  *		before throttling back.
182  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
183  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
184  * INTERACT_THRESH:	Threshold for placement on the current runq.
185  */
186 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
187 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
188 #define	SCHED_INTERACT_MAX	(100)
189 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
190 #define	SCHED_INTERACT_THRESH	(30)
191 
192 /*
193  * These parameters determine the slice behavior for batch work.
194  */
195 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
196 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
197 
198 /* Flags kept in td_flags. */
199 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
200 
201 /*
202  * tickincr:		Converts a stathz tick into a hz domain scaled by
203  *			the shift factor.  Without the shift the error rate
204  *			due to rounding would be unacceptably high.
205  * realstathz:		stathz is sometimes 0 and run off of hz.
206  * sched_slice:		Runtime of each thread before rescheduling.
207  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
208  */
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127;	/* reset during boot. */
212 static int sched_slice = 10;	/* reset during boot. */
213 static int sched_slice_min = 1;	/* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int preempt_thresh = 0;
222 #endif
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
226 
227 /*
228  * tdq - per processor runqs and statistics.  All fields are protected by the
229  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
230  * locking in sched_pickcpu();
231  */
232 struct tdq {
233 	/*
234 	 * Ordered to improve efficiency of cpu_search() and switch().
235 	 * tdq_lock is padded to avoid false sharing with tdq_load and
236 	 * tdq_cpu_idle.
237 	 */
238 	struct mtx_padalign tdq_lock;		/* run queue lock. */
239 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
240 	volatile int	tdq_load;		/* Aggregate load. */
241 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
242 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
243 	volatile int	tdq_transferable;	/* Transferable thread count. */
244 	volatile short	tdq_switchcnt;		/* Switches this tick. */
245 	volatile short	tdq_oldswitchcnt;	/* Switches last tick. */
246 	u_char		tdq_lowpri;		/* Lowest priority thread. */
247 	u_char		tdq_ipipending;		/* IPI pending. */
248 	u_char		tdq_idx;		/* Current insert index. */
249 	u_char		tdq_ridx;		/* Current removal index. */
250 	int		tdq_id;			/* cpuid. */
251 	struct runq	tdq_realtime;		/* real-time run queue. */
252 	struct runq	tdq_timeshare;		/* timeshare run queue. */
253 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
254 	char		tdq_name[TDQ_NAME_LEN];
255 #ifdef KTR
256 	char		tdq_loadname[TDQ_LOADNAME_LEN];
257 #endif
258 } __aligned(64);
259 
260 /* Idle thread states and config. */
261 #define	TDQ_RUNNING	1
262 #define	TDQ_IDLE	2
263 
264 #ifdef SMP
265 struct cpu_group *cpu_top;		/* CPU topology */
266 
267 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
268 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
269 
270 /*
271  * Run-time tunables.
272  */
273 static int rebalance = 1;
274 static int balance_interval = 128;	/* Default set in sched_initticks(). */
275 static int affinity;
276 static int steal_idle = 1;
277 static int steal_thresh = 2;
278 static int always_steal = 0;
279 static int trysteal_limit = 2;
280 
281 /*
282  * One thread queue per processor.
283  */
284 static struct tdq	*balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(struct tdq, tdq);
287 DPCPU_DEFINE_STATIC(uint32_t, randomval);
288 
289 #define	TDQ_SELF()	((struct tdq *)PCPU_GET(sched))
290 #define	TDQ_CPU(x)	(DPCPU_ID_PTR((x), tdq))
291 #define	TDQ_ID(x)	((x)->tdq_id)
292 #else	/* !SMP */
293 static struct tdq	tdq_cpu;
294 
295 #define	TDQ_ID(x)	(0)
296 #define	TDQ_SELF()	(&tdq_cpu)
297 #define	TDQ_CPU(x)	(&tdq_cpu)
298 #endif
299 
300 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
301 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
302 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
303 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
304 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
305 
306 static void sched_priority(struct thread *);
307 static void sched_thread_priority(struct thread *, u_char);
308 static int sched_interact_score(struct thread *);
309 static void sched_interact_update(struct thread *);
310 static void sched_interact_fork(struct thread *);
311 static void sched_pctcpu_update(struct td_sched *, int);
312 
313 /* Operations on per processor queues */
314 static struct thread *tdq_choose(struct tdq *);
315 static void tdq_setup(struct tdq *, int i);
316 static void tdq_load_add(struct tdq *, struct thread *);
317 static void tdq_load_rem(struct tdq *, struct thread *);
318 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
319 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
320 static inline int sched_shouldpreempt(int, int, int);
321 void tdq_print(int cpu);
322 static void runq_print(struct runq *rq);
323 static void tdq_add(struct tdq *, struct thread *, int);
324 #ifdef SMP
325 static struct thread *tdq_move(struct tdq *, struct tdq *);
326 static int tdq_idled(struct tdq *);
327 static void tdq_notify(struct tdq *, struct thread *);
328 static struct thread *tdq_steal(struct tdq *, int);
329 static struct thread *runq_steal(struct runq *, int);
330 static int sched_pickcpu(struct thread *, int);
331 static void sched_balance(void);
332 static int sched_balance_pair(struct tdq *, struct tdq *);
333 static inline struct tdq *sched_setcpu(struct thread *, int, int);
334 static inline void thread_unblock_switch(struct thread *, struct mtx *);
335 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
336 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
337 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
338     struct cpu_group *cg, int indent);
339 #endif
340 
341 static void sched_setup(void *dummy);
342 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
343 
344 static void sched_initticks(void *dummy);
345 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
346     NULL);
347 
348 SDT_PROVIDER_DEFINE(sched);
349 
350 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
351     "struct proc *", "uint8_t");
352 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
353     "struct proc *", "void *");
354 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
355     "struct proc *", "void *", "int");
356 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
357     "struct proc *", "uint8_t", "struct thread *");
358 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
359 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
360     "struct proc *");
361 SDT_PROBE_DEFINE(sched, , , on__cpu);
362 SDT_PROBE_DEFINE(sched, , , remain__cpu);
363 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
364     "struct proc *");
365 
366 /*
367  * Print the threads waiting on a run-queue.
368  */
369 static void
370 runq_print(struct runq *rq)
371 {
372 	struct rqhead *rqh;
373 	struct thread *td;
374 	int pri;
375 	int j;
376 	int i;
377 
378 	for (i = 0; i < RQB_LEN; i++) {
379 		printf("\t\trunq bits %d 0x%zx\n",
380 		    i, rq->rq_status.rqb_bits[i]);
381 		for (j = 0; j < RQB_BPW; j++)
382 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
383 				pri = j + (i << RQB_L2BPW);
384 				rqh = &rq->rq_queues[pri];
385 				TAILQ_FOREACH(td, rqh, td_runq) {
386 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
387 					    td, td->td_name, td->td_priority,
388 					    td->td_rqindex, pri);
389 				}
390 			}
391 	}
392 }
393 
394 /*
395  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
396  */
397 void
398 tdq_print(int cpu)
399 {
400 	struct tdq *tdq;
401 
402 	tdq = TDQ_CPU(cpu);
403 
404 	printf("tdq %d:\n", TDQ_ID(tdq));
405 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
406 	printf("\tLock name:      %s\n", tdq->tdq_name);
407 	printf("\tload:           %d\n", tdq->tdq_load);
408 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
409 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
410 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
411 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
412 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
413 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
414 	printf("\trealtime runq:\n");
415 	runq_print(&tdq->tdq_realtime);
416 	printf("\ttimeshare runq:\n");
417 	runq_print(&tdq->tdq_timeshare);
418 	printf("\tidle runq:\n");
419 	runq_print(&tdq->tdq_idle);
420 }
421 
422 static inline int
423 sched_shouldpreempt(int pri, int cpri, int remote)
424 {
425 	/*
426 	 * If the new priority is not better than the current priority there is
427 	 * nothing to do.
428 	 */
429 	if (pri >= cpri)
430 		return (0);
431 	/*
432 	 * Always preempt idle.
433 	 */
434 	if (cpri >= PRI_MIN_IDLE)
435 		return (1);
436 	/*
437 	 * If preemption is disabled don't preempt others.
438 	 */
439 	if (preempt_thresh == 0)
440 		return (0);
441 	/*
442 	 * Preempt if we exceed the threshold.
443 	 */
444 	if (pri <= preempt_thresh)
445 		return (1);
446 	/*
447 	 * If we're interactive or better and there is non-interactive
448 	 * or worse running preempt only remote processors.
449 	 */
450 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
451 		return (1);
452 	return (0);
453 }
454 
455 /*
456  * Add a thread to the actual run-queue.  Keeps transferable counts up to
457  * date with what is actually on the run-queue.  Selects the correct
458  * queue position for timeshare threads.
459  */
460 static __inline void
461 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
462 {
463 	struct td_sched *ts;
464 	u_char pri;
465 
466 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
467 	THREAD_LOCK_ASSERT(td, MA_OWNED);
468 
469 	pri = td->td_priority;
470 	ts = td_get_sched(td);
471 	TD_SET_RUNQ(td);
472 	if (THREAD_CAN_MIGRATE(td)) {
473 		tdq->tdq_transferable++;
474 		ts->ts_flags |= TSF_XFERABLE;
475 	}
476 	if (pri < PRI_MIN_BATCH) {
477 		ts->ts_runq = &tdq->tdq_realtime;
478 	} else if (pri <= PRI_MAX_BATCH) {
479 		ts->ts_runq = &tdq->tdq_timeshare;
480 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
481 			("Invalid priority %d on timeshare runq", pri));
482 		/*
483 		 * This queue contains only priorities between MIN and MAX
484 		 * realtime.  Use the whole queue to represent these values.
485 		 */
486 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
487 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
488 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
489 			/*
490 			 * This effectively shortens the queue by one so we
491 			 * can have a one slot difference between idx and
492 			 * ridx while we wait for threads to drain.
493 			 */
494 			if (tdq->tdq_ridx != tdq->tdq_idx &&
495 			    pri == tdq->tdq_ridx)
496 				pri = (unsigned char)(pri - 1) % RQ_NQS;
497 		} else
498 			pri = tdq->tdq_ridx;
499 		runq_add_pri(ts->ts_runq, td, pri, flags);
500 		return;
501 	} else
502 		ts->ts_runq = &tdq->tdq_idle;
503 	runq_add(ts->ts_runq, td, flags);
504 }
505 
506 /*
507  * Remove a thread from a run-queue.  This typically happens when a thread
508  * is selected to run.  Running threads are not on the queue and the
509  * transferable count does not reflect them.
510  */
511 static __inline void
512 tdq_runq_rem(struct tdq *tdq, struct thread *td)
513 {
514 	struct td_sched *ts;
515 
516 	ts = td_get_sched(td);
517 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
518 	KASSERT(ts->ts_runq != NULL,
519 	    ("tdq_runq_remove: thread %p null ts_runq", td));
520 	if (ts->ts_flags & TSF_XFERABLE) {
521 		tdq->tdq_transferable--;
522 		ts->ts_flags &= ~TSF_XFERABLE;
523 	}
524 	if (ts->ts_runq == &tdq->tdq_timeshare) {
525 		if (tdq->tdq_idx != tdq->tdq_ridx)
526 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
527 		else
528 			runq_remove_idx(ts->ts_runq, td, NULL);
529 	} else
530 		runq_remove(ts->ts_runq, td);
531 }
532 
533 /*
534  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
535  * for this thread to the referenced thread queue.
536  */
537 static void
538 tdq_load_add(struct tdq *tdq, struct thread *td)
539 {
540 
541 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
542 	THREAD_LOCK_ASSERT(td, MA_OWNED);
543 
544 	tdq->tdq_load++;
545 	if ((td->td_flags & TDF_NOLOAD) == 0)
546 		tdq->tdq_sysload++;
547 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
548 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
549 }
550 
551 /*
552  * Remove the load from a thread that is transitioning to a sleep state or
553  * exiting.
554  */
555 static void
556 tdq_load_rem(struct tdq *tdq, struct thread *td)
557 {
558 
559 	THREAD_LOCK_ASSERT(td, MA_OWNED);
560 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
561 	KASSERT(tdq->tdq_load != 0,
562 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
563 
564 	tdq->tdq_load--;
565 	if ((td->td_flags & TDF_NOLOAD) == 0)
566 		tdq->tdq_sysload--;
567 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
568 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 }
570 
571 /*
572  * Bound timeshare latency by decreasing slice size as load increases.  We
573  * consider the maximum latency as the sum of the threads waiting to run
574  * aside from curthread and target no more than sched_slice latency but
575  * no less than sched_slice_min runtime.
576  */
577 static inline int
578 tdq_slice(struct tdq *tdq)
579 {
580 	int load;
581 
582 	/*
583 	 * It is safe to use sys_load here because this is called from
584 	 * contexts where timeshare threads are running and so there
585 	 * cannot be higher priority load in the system.
586 	 */
587 	load = tdq->tdq_sysload - 1;
588 	if (load >= SCHED_SLICE_MIN_DIVISOR)
589 		return (sched_slice_min);
590 	if (load <= 1)
591 		return (sched_slice);
592 	return (sched_slice / load);
593 }
594 
595 /*
596  * Set lowpri to its exact value by searching the run-queue and
597  * evaluating curthread.  curthread may be passed as an optimization.
598  */
599 static void
600 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
601 {
602 	struct thread *td;
603 
604 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
605 	if (ctd == NULL)
606 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
607 	td = tdq_choose(tdq);
608 	if (td == NULL || td->td_priority > ctd->td_priority)
609 		tdq->tdq_lowpri = ctd->td_priority;
610 	else
611 		tdq->tdq_lowpri = td->td_priority;
612 }
613 
614 #ifdef SMP
615 /*
616  * We need some randomness. Implement a classic Linear Congruential
617  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
618  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
619  * of the random state (in the low bits of our answer) to keep
620  * the maximum randomness.
621  */
622 static uint32_t
623 sched_random(void)
624 {
625 	uint32_t *rndptr;
626 
627 	rndptr = DPCPU_PTR(randomval);
628 	*rndptr = *rndptr * 69069 + 5;
629 
630 	return (*rndptr >> 16);
631 }
632 
633 struct cpu_search {
634 	cpuset_t cs_mask;
635 	u_int	cs_prefer;
636 	int	cs_pri;		/* Min priority for low. */
637 	int	cs_limit;	/* Max load for low, min load for high. */
638 	int	cs_cpu;
639 	int	cs_load;
640 };
641 
642 #define	CPU_SEARCH_LOWEST	0x1
643 #define	CPU_SEARCH_HIGHEST	0x2
644 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
645 
646 #define	CPUSET_FOREACH(cpu, mask)				\
647 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
648 		if (CPU_ISSET(cpu, &mask))
649 
650 static __always_inline int cpu_search(const struct cpu_group *cg,
651     struct cpu_search *low, struct cpu_search *high, const int match);
652 int __noinline cpu_search_lowest(const struct cpu_group *cg,
653     struct cpu_search *low);
654 int __noinline cpu_search_highest(const struct cpu_group *cg,
655     struct cpu_search *high);
656 int __noinline cpu_search_both(const struct cpu_group *cg,
657     struct cpu_search *low, struct cpu_search *high);
658 
659 /*
660  * Search the tree of cpu_groups for the lowest or highest loaded cpu
661  * according to the match argument.  This routine actually compares the
662  * load on all paths through the tree and finds the least loaded cpu on
663  * the least loaded path, which may differ from the least loaded cpu in
664  * the system.  This balances work among caches and buses.
665  *
666  * This inline is instantiated in three forms below using constants for the
667  * match argument.  It is reduced to the minimum set for each case.  It is
668  * also recursive to the depth of the tree.
669  */
670 static __always_inline int
671 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
672     struct cpu_search *high, const int match)
673 {
674 	struct cpu_search lgroup;
675 	struct cpu_search hgroup;
676 	cpuset_t cpumask;
677 	struct cpu_group *child;
678 	struct tdq *tdq;
679 	int cpu, i, hload, lload, load, total, rnd;
680 
681 	total = 0;
682 	cpumask = cg->cg_mask;
683 	if (match & CPU_SEARCH_LOWEST) {
684 		lload = INT_MAX;
685 		lgroup = *low;
686 	}
687 	if (match & CPU_SEARCH_HIGHEST) {
688 		hload = INT_MIN;
689 		hgroup = *high;
690 	}
691 
692 	/* Iterate through the child CPU groups and then remaining CPUs. */
693 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
694 		if (i == 0) {
695 #ifdef HAVE_INLINE_FFSL
696 			cpu = CPU_FFS(&cpumask) - 1;
697 #else
698 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
699 				cpu--;
700 #endif
701 			if (cpu < 0)
702 				break;
703 			child = NULL;
704 		} else
705 			child = &cg->cg_child[i - 1];
706 
707 		if (match & CPU_SEARCH_LOWEST)
708 			lgroup.cs_cpu = -1;
709 		if (match & CPU_SEARCH_HIGHEST)
710 			hgroup.cs_cpu = -1;
711 		if (child) {			/* Handle child CPU group. */
712 			CPU_NAND(&cpumask, &child->cg_mask);
713 			switch (match) {
714 			case CPU_SEARCH_LOWEST:
715 				load = cpu_search_lowest(child, &lgroup);
716 				break;
717 			case CPU_SEARCH_HIGHEST:
718 				load = cpu_search_highest(child, &hgroup);
719 				break;
720 			case CPU_SEARCH_BOTH:
721 				load = cpu_search_both(child, &lgroup, &hgroup);
722 				break;
723 			}
724 		} else {			/* Handle child CPU. */
725 			CPU_CLR(cpu, &cpumask);
726 			tdq = TDQ_CPU(cpu);
727 			load = tdq->tdq_load * 256;
728 			rnd = sched_random() % 32;
729 			if (match & CPU_SEARCH_LOWEST) {
730 				if (cpu == low->cs_prefer)
731 					load -= 64;
732 				/* If that CPU is allowed and get data. */
733 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
734 				    tdq->tdq_load <= lgroup.cs_limit &&
735 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
736 					lgroup.cs_cpu = cpu;
737 					lgroup.cs_load = load - rnd;
738 				}
739 			}
740 			if (match & CPU_SEARCH_HIGHEST)
741 				if (tdq->tdq_load >= hgroup.cs_limit &&
742 				    tdq->tdq_transferable &&
743 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
744 					hgroup.cs_cpu = cpu;
745 					hgroup.cs_load = load - rnd;
746 				}
747 		}
748 		total += load;
749 
750 		/* We have info about child item. Compare it. */
751 		if (match & CPU_SEARCH_LOWEST) {
752 			if (lgroup.cs_cpu >= 0 &&
753 			    (load < lload ||
754 			     (load == lload && lgroup.cs_load < low->cs_load))) {
755 				lload = load;
756 				low->cs_cpu = lgroup.cs_cpu;
757 				low->cs_load = lgroup.cs_load;
758 			}
759 		}
760 		if (match & CPU_SEARCH_HIGHEST)
761 			if (hgroup.cs_cpu >= 0 &&
762 			    (load > hload ||
763 			     (load == hload && hgroup.cs_load > high->cs_load))) {
764 				hload = load;
765 				high->cs_cpu = hgroup.cs_cpu;
766 				high->cs_load = hgroup.cs_load;
767 			}
768 		if (child) {
769 			i--;
770 			if (i == 0 && CPU_EMPTY(&cpumask))
771 				break;
772 		}
773 #ifndef HAVE_INLINE_FFSL
774 		else
775 			cpu--;
776 #endif
777 	}
778 	return (total);
779 }
780 
781 /*
782  * cpu_search instantiations must pass constants to maintain the inline
783  * optimization.
784  */
785 int
786 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
787 {
788 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
789 }
790 
791 int
792 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
793 {
794 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
795 }
796 
797 int
798 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
799     struct cpu_search *high)
800 {
801 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
802 }
803 
804 /*
805  * Find the cpu with the least load via the least loaded path that has a
806  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
807  * acceptable.
808  */
809 static inline int
810 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
811     int prefer)
812 {
813 	struct cpu_search low;
814 
815 	low.cs_cpu = -1;
816 	low.cs_prefer = prefer;
817 	low.cs_mask = mask;
818 	low.cs_pri = pri;
819 	low.cs_limit = maxload;
820 	cpu_search_lowest(cg, &low);
821 	return low.cs_cpu;
822 }
823 
824 /*
825  * Find the cpu with the highest load via the highest loaded path.
826  */
827 static inline int
828 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
829 {
830 	struct cpu_search high;
831 
832 	high.cs_cpu = -1;
833 	high.cs_mask = mask;
834 	high.cs_limit = minload;
835 	cpu_search_highest(cg, &high);
836 	return high.cs_cpu;
837 }
838 
839 static void
840 sched_balance_group(struct cpu_group *cg)
841 {
842 	struct tdq *tdq;
843 	cpuset_t hmask, lmask;
844 	int high, low, anylow;
845 
846 	CPU_FILL(&hmask);
847 	for (;;) {
848 		high = sched_highest(cg, hmask, 2);
849 		/* Stop if there is no more CPU with transferrable threads. */
850 		if (high == -1)
851 			break;
852 		CPU_CLR(high, &hmask);
853 		CPU_COPY(&hmask, &lmask);
854 		/* Stop if there is no more CPU left for low. */
855 		if (CPU_EMPTY(&lmask))
856 			break;
857 		anylow = 1;
858 		tdq = TDQ_CPU(high);
859 nextlow:
860 		low = sched_lowest(cg, lmask, -1, tdq->tdq_load - 1, high);
861 		/* Stop if we looked well and found no less loaded CPU. */
862 		if (anylow && low == -1)
863 			break;
864 		/* Go to next high if we found no less loaded CPU. */
865 		if (low == -1)
866 			continue;
867 		/* Transfer thread from high to low. */
868 		if (sched_balance_pair(tdq, TDQ_CPU(low))) {
869 			/* CPU that got thread can no longer be a donor. */
870 			CPU_CLR(low, &hmask);
871 		} else {
872 			/*
873 			 * If failed, then there is no threads on high
874 			 * that can run on this low. Drop low from low
875 			 * mask and look for different one.
876 			 */
877 			CPU_CLR(low, &lmask);
878 			anylow = 0;
879 			goto nextlow;
880 		}
881 	}
882 }
883 
884 static void
885 sched_balance(void)
886 {
887 	struct tdq *tdq;
888 
889 	balance_ticks = max(balance_interval / 2, 1) +
890 	    (sched_random() % balance_interval);
891 	tdq = TDQ_SELF();
892 	TDQ_UNLOCK(tdq);
893 	sched_balance_group(cpu_top);
894 	TDQ_LOCK(tdq);
895 }
896 
897 /*
898  * Lock two thread queues using their address to maintain lock order.
899  */
900 static void
901 tdq_lock_pair(struct tdq *one, struct tdq *two)
902 {
903 	if (one < two) {
904 		TDQ_LOCK(one);
905 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
906 	} else {
907 		TDQ_LOCK(two);
908 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
909 	}
910 }
911 
912 /*
913  * Unlock two thread queues.  Order is not important here.
914  */
915 static void
916 tdq_unlock_pair(struct tdq *one, struct tdq *two)
917 {
918 	TDQ_UNLOCK(one);
919 	TDQ_UNLOCK(two);
920 }
921 
922 /*
923  * Transfer load between two imbalanced thread queues.
924  */
925 static int
926 sched_balance_pair(struct tdq *high, struct tdq *low)
927 {
928 	struct thread *td;
929 	int cpu;
930 
931 	tdq_lock_pair(high, low);
932 	td = NULL;
933 	/*
934 	 * Transfer a thread from high to low.
935 	 */
936 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
937 	    (td = tdq_move(high, low)) != NULL) {
938 		/*
939 		 * In case the target isn't the current cpu notify it of the
940 		 * new load, possibly sending an IPI to force it to reschedule.
941 		 */
942 		cpu = TDQ_ID(low);
943 		if (cpu != PCPU_GET(cpuid))
944 			tdq_notify(low, td);
945 	}
946 	tdq_unlock_pair(high, low);
947 	return (td != NULL);
948 }
949 
950 /*
951  * Move a thread from one thread queue to another.
952  */
953 static struct thread *
954 tdq_move(struct tdq *from, struct tdq *to)
955 {
956 	struct td_sched *ts;
957 	struct thread *td;
958 	struct tdq *tdq;
959 	int cpu;
960 
961 	TDQ_LOCK_ASSERT(from, MA_OWNED);
962 	TDQ_LOCK_ASSERT(to, MA_OWNED);
963 
964 	tdq = from;
965 	cpu = TDQ_ID(to);
966 	td = tdq_steal(tdq, cpu);
967 	if (td == NULL)
968 		return (NULL);
969 	ts = td_get_sched(td);
970 	/*
971 	 * Although the run queue is locked the thread may be blocked.  Lock
972 	 * it to clear this and acquire the run-queue lock.
973 	 */
974 	thread_lock(td);
975 	/* Drop recursive lock on from acquired via thread_lock(). */
976 	TDQ_UNLOCK(from);
977 	sched_rem(td);
978 	ts->ts_cpu = cpu;
979 	td->td_lock = TDQ_LOCKPTR(to);
980 	tdq_add(to, td, SRQ_YIELDING);
981 	return (td);
982 }
983 
984 /*
985  * This tdq has idled.  Try to steal a thread from another cpu and switch
986  * to it.
987  */
988 static int
989 tdq_idled(struct tdq *tdq)
990 {
991 	struct cpu_group *cg;
992 	struct tdq *steal;
993 	cpuset_t mask;
994 	int cpu, switchcnt;
995 
996 	if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
997 		return (1);
998 	CPU_FILL(&mask);
999 	CPU_CLR(PCPU_GET(cpuid), &mask);
1000     restart:
1001 	switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
1002 	for (cg = tdq->tdq_cg; ; ) {
1003 		cpu = sched_highest(cg, mask, steal_thresh);
1004 		/*
1005 		 * We were assigned a thread but not preempted.  Returning
1006 		 * 0 here will cause our caller to switch to it.
1007 		 */
1008 		if (tdq->tdq_load)
1009 			return (0);
1010 		if (cpu == -1) {
1011 			cg = cg->cg_parent;
1012 			if (cg == NULL)
1013 				return (1);
1014 			continue;
1015 		}
1016 		steal = TDQ_CPU(cpu);
1017 		/*
1018 		 * The data returned by sched_highest() is stale and
1019 		 * the chosen CPU no longer has an eligible thread.
1020 		 *
1021 		 * Testing this ahead of tdq_lock_pair() only catches
1022 		 * this situation about 20% of the time on an 8 core
1023 		 * 16 thread Ryzen 7, but it still helps performance.
1024 		 */
1025 		if (steal->tdq_load < steal_thresh ||
1026 		    steal->tdq_transferable == 0)
1027 			goto restart;
1028 		tdq_lock_pair(tdq, steal);
1029 		/*
1030 		 * We were assigned a thread while waiting for the locks.
1031 		 * Switch to it now instead of stealing a thread.
1032 		 */
1033 		if (tdq->tdq_load)
1034 			break;
1035 		/*
1036 		 * The data returned by sched_highest() is stale and
1037 		 * the chosen CPU no longer has an eligible thread, or
1038 		 * we were preempted and the CPU loading info may be out
1039 		 * of date.  The latter is rare.  In either case restart
1040 		 * the search.
1041 		 */
1042 		if (steal->tdq_load < steal_thresh ||
1043 		    steal->tdq_transferable == 0 ||
1044 		    switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1045 			tdq_unlock_pair(tdq, steal);
1046 			goto restart;
1047 		}
1048 		/*
1049 		 * Steal the thread and switch to it.
1050 		 */
1051 		if (tdq_move(steal, tdq) != NULL)
1052 			break;
1053 		/*
1054 		 * We failed to acquire a thread even though it looked
1055 		 * like one was available.  This could be due to affinity
1056 		 * restrictions or for other reasons.  Loop again after
1057 		 * removing this CPU from the set.  The restart logic
1058 		 * above does not restore this CPU to the set due to the
1059 		 * likelyhood of failing here again.
1060 		 */
1061 		CPU_CLR(cpu, &mask);
1062 		tdq_unlock_pair(tdq, steal);
1063 	}
1064 	TDQ_UNLOCK(steal);
1065 	mi_switch(SW_VOL | SWT_IDLE, NULL);
1066 	thread_unlock(curthread);
1067 	return (0);
1068 }
1069 
1070 /*
1071  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1072  */
1073 static void
1074 tdq_notify(struct tdq *tdq, struct thread *td)
1075 {
1076 	struct thread *ctd;
1077 	int pri;
1078 	int cpu;
1079 
1080 	if (tdq->tdq_ipipending)
1081 		return;
1082 	cpu = td_get_sched(td)->ts_cpu;
1083 	pri = td->td_priority;
1084 	ctd = pcpu_find(cpu)->pc_curthread;
1085 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1086 		return;
1087 
1088 	/*
1089 	 * Make sure that our caller's earlier update to tdq_load is
1090 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1091 	 * accesses both of them without locks, and the order is important.
1092 	 */
1093 	atomic_thread_fence_seq_cst();
1094 
1095 	if (TD_IS_IDLETHREAD(ctd)) {
1096 		/*
1097 		 * If the MD code has an idle wakeup routine try that before
1098 		 * falling back to IPI.
1099 		 */
1100 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1101 			return;
1102 	}
1103 	tdq->tdq_ipipending = 1;
1104 	ipi_cpu(cpu, IPI_PREEMPT);
1105 }
1106 
1107 /*
1108  * Steals load from a timeshare queue.  Honors the rotating queue head
1109  * index.
1110  */
1111 static struct thread *
1112 runq_steal_from(struct runq *rq, int cpu, u_char start)
1113 {
1114 	struct rqbits *rqb;
1115 	struct rqhead *rqh;
1116 	struct thread *td, *first;
1117 	int bit;
1118 	int i;
1119 
1120 	rqb = &rq->rq_status;
1121 	bit = start & (RQB_BPW -1);
1122 	first = NULL;
1123 again:
1124 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1125 		if (rqb->rqb_bits[i] == 0)
1126 			continue;
1127 		if (bit == 0)
1128 			bit = RQB_FFS(rqb->rqb_bits[i]);
1129 		for (; bit < RQB_BPW; bit++) {
1130 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1131 				continue;
1132 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1133 			TAILQ_FOREACH(td, rqh, td_runq) {
1134 				if (first && THREAD_CAN_MIGRATE(td) &&
1135 				    THREAD_CAN_SCHED(td, cpu))
1136 					return (td);
1137 				first = td;
1138 			}
1139 		}
1140 	}
1141 	if (start != 0) {
1142 		start = 0;
1143 		goto again;
1144 	}
1145 
1146 	if (first && THREAD_CAN_MIGRATE(first) &&
1147 	    THREAD_CAN_SCHED(first, cpu))
1148 		return (first);
1149 	return (NULL);
1150 }
1151 
1152 /*
1153  * Steals load from a standard linear queue.
1154  */
1155 static struct thread *
1156 runq_steal(struct runq *rq, int cpu)
1157 {
1158 	struct rqhead *rqh;
1159 	struct rqbits *rqb;
1160 	struct thread *td;
1161 	int word;
1162 	int bit;
1163 
1164 	rqb = &rq->rq_status;
1165 	for (word = 0; word < RQB_LEN; word++) {
1166 		if (rqb->rqb_bits[word] == 0)
1167 			continue;
1168 		for (bit = 0; bit < RQB_BPW; bit++) {
1169 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1170 				continue;
1171 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1172 			TAILQ_FOREACH(td, rqh, td_runq)
1173 				if (THREAD_CAN_MIGRATE(td) &&
1174 				    THREAD_CAN_SCHED(td, cpu))
1175 					return (td);
1176 		}
1177 	}
1178 	return (NULL);
1179 }
1180 
1181 /*
1182  * Attempt to steal a thread in priority order from a thread queue.
1183  */
1184 static struct thread *
1185 tdq_steal(struct tdq *tdq, int cpu)
1186 {
1187 	struct thread *td;
1188 
1189 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1190 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1191 		return (td);
1192 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1193 	    cpu, tdq->tdq_ridx)) != NULL)
1194 		return (td);
1195 	return (runq_steal(&tdq->tdq_idle, cpu));
1196 }
1197 
1198 /*
1199  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1200  * current lock and returns with the assigned queue locked.
1201  */
1202 static inline struct tdq *
1203 sched_setcpu(struct thread *td, int cpu, int flags)
1204 {
1205 
1206 	struct tdq *tdq;
1207 
1208 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1209 	tdq = TDQ_CPU(cpu);
1210 	td_get_sched(td)->ts_cpu = cpu;
1211 	/*
1212 	 * If the lock matches just return the queue.
1213 	 */
1214 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1215 		return (tdq);
1216 #ifdef notyet
1217 	/*
1218 	 * If the thread isn't running its lockptr is a
1219 	 * turnstile or a sleepqueue.  We can just lock_set without
1220 	 * blocking.
1221 	 */
1222 	if (TD_CAN_RUN(td)) {
1223 		TDQ_LOCK(tdq);
1224 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1225 		return (tdq);
1226 	}
1227 #endif
1228 	/*
1229 	 * The hard case, migration, we need to block the thread first to
1230 	 * prevent order reversals with other cpus locks.
1231 	 */
1232 	spinlock_enter();
1233 	thread_lock_block(td);
1234 	TDQ_LOCK(tdq);
1235 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1236 	spinlock_exit();
1237 	return (tdq);
1238 }
1239 
1240 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1241 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1242 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1243 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1244 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1245 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1246 
1247 static int
1248 sched_pickcpu(struct thread *td, int flags)
1249 {
1250 	struct cpu_group *cg, *ccg;
1251 	struct td_sched *ts;
1252 	struct tdq *tdq;
1253 	cpuset_t mask;
1254 	int cpu, pri, self, intr;
1255 
1256 	self = PCPU_GET(cpuid);
1257 	ts = td_get_sched(td);
1258 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1259 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1260 	if (smp_started == 0)
1261 		return (self);
1262 	/*
1263 	 * Don't migrate a running thread from sched_switch().
1264 	 */
1265 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1266 		return (ts->ts_cpu);
1267 	/*
1268 	 * Prefer to run interrupt threads on the processors that generate
1269 	 * the interrupt.
1270 	 */
1271 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1272 	    curthread->td_intr_nesting_level) {
1273 		ts->ts_cpu = self;
1274 		intr = 1;
1275 	} else
1276 		intr = 0;
1277 	/*
1278 	 * If the thread can run on the last cpu and the affinity has not
1279 	 * expired and it is idle, run it there.
1280 	 */
1281 	tdq = TDQ_CPU(ts->ts_cpu);
1282 	cg = tdq->tdq_cg;
1283 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1284 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1285 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1286 		if (!intr && cg->cg_flags & CG_FLAG_THREAD) {
1287 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1288 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1289 					break;
1290 			}
1291 		} else
1292 			cpu = INT_MAX;
1293 		if (cpu > mp_maxid) {
1294 			SCHED_STAT_INC(pickcpu_idle_affinity);
1295 			return (ts->ts_cpu);
1296 		}
1297 	}
1298 	/*
1299 	 * Search for the last level cache CPU group in the tree.
1300 	 * Skip SMT, identical groups and caches with expired affinity.
1301 	 * Interrupt threads affinity is explicit and never expires.
1302 	 */
1303 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1304 		if (cg->cg_flags & CG_FLAG_THREAD)
1305 			continue;
1306 		if (cg->cg_children == 1 || cg->cg_count == 1)
1307 			continue;
1308 		if (cg->cg_level == CG_SHARE_NONE ||
1309 		    (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1310 			continue;
1311 		ccg = cg;
1312 	}
1313 	/* Found LLC shared by all CPUs, so do a global search. */
1314 	if (ccg == cpu_top)
1315 		ccg = NULL;
1316 	cpu = -1;
1317 	mask = td->td_cpuset->cs_mask;
1318 	pri = td->td_priority;
1319 	/*
1320 	 * Try hard to keep interrupts within found LLC.  Search the LLC for
1321 	 * the least loaded CPU we can run now.  For NUMA systems it should
1322 	 * be within target domain, and it also reduces scheduling overhead.
1323 	 */
1324 	if (ccg != NULL && intr) {
1325 		cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1326 		if (cpu >= 0)
1327 			SCHED_STAT_INC(pickcpu_intrbind);
1328 	} else
1329 	/* Search the LLC for the least loaded idle CPU we can run now. */
1330 	if (ccg != NULL) {
1331 		cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1332 		    INT_MAX, ts->ts_cpu);
1333 		if (cpu >= 0)
1334 			SCHED_STAT_INC(pickcpu_affinity);
1335 	}
1336 	/* Search globally for the least loaded CPU we can run now. */
1337 	if (cpu < 0) {
1338 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1339 		if (cpu >= 0)
1340 			SCHED_STAT_INC(pickcpu_lowest);
1341 	}
1342 	/* Search globally for the least loaded CPU. */
1343 	if (cpu < 0) {
1344 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1345 		if (cpu >= 0)
1346 			SCHED_STAT_INC(pickcpu_lowest);
1347 	}
1348 	KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1349 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1350 	/*
1351 	 * Compare the lowest loaded cpu to current cpu.
1352 	 */
1353 	tdq = TDQ_CPU(cpu);
1354 	if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1355 	    tdq->tdq_lowpri < PRI_MIN_IDLE &&
1356 	    TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1357 		SCHED_STAT_INC(pickcpu_local);
1358 		cpu = self;
1359 	}
1360 	if (cpu != ts->ts_cpu)
1361 		SCHED_STAT_INC(pickcpu_migration);
1362 	return (cpu);
1363 }
1364 #endif
1365 
1366 /*
1367  * Pick the highest priority task we have and return it.
1368  */
1369 static struct thread *
1370 tdq_choose(struct tdq *tdq)
1371 {
1372 	struct thread *td;
1373 
1374 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1375 	td = runq_choose(&tdq->tdq_realtime);
1376 	if (td != NULL)
1377 		return (td);
1378 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1379 	if (td != NULL) {
1380 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1381 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1382 		    td->td_priority));
1383 		return (td);
1384 	}
1385 	td = runq_choose(&tdq->tdq_idle);
1386 	if (td != NULL) {
1387 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1388 		    ("tdq_choose: Invalid priority on idle queue %d",
1389 		    td->td_priority));
1390 		return (td);
1391 	}
1392 
1393 	return (NULL);
1394 }
1395 
1396 /*
1397  * Initialize a thread queue.
1398  */
1399 static void
1400 tdq_setup(struct tdq *tdq, int id)
1401 {
1402 
1403 	if (bootverbose)
1404 		printf("ULE: setup cpu %d\n", id);
1405 	runq_init(&tdq->tdq_realtime);
1406 	runq_init(&tdq->tdq_timeshare);
1407 	runq_init(&tdq->tdq_idle);
1408 	tdq->tdq_id = id;
1409 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1410 	    "sched lock %d", (int)TDQ_ID(tdq));
1411 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1412 	    MTX_SPIN | MTX_RECURSE);
1413 #ifdef KTR
1414 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1415 	    "CPU %d load", (int)TDQ_ID(tdq));
1416 #endif
1417 }
1418 
1419 #ifdef SMP
1420 static void
1421 sched_setup_smp(void)
1422 {
1423 	struct tdq *tdq;
1424 	int i;
1425 
1426 	cpu_top = smp_topo();
1427 	CPU_FOREACH(i) {
1428 		tdq = DPCPU_ID_PTR(i, tdq);
1429 		tdq_setup(tdq, i);
1430 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1431 		if (tdq->tdq_cg == NULL)
1432 			panic("Can't find cpu group for %d\n", i);
1433 	}
1434 	PCPU_SET(sched, DPCPU_PTR(tdq));
1435 	balance_tdq = TDQ_SELF();
1436 }
1437 #endif
1438 
1439 /*
1440  * Setup the thread queues and initialize the topology based on MD
1441  * information.
1442  */
1443 static void
1444 sched_setup(void *dummy)
1445 {
1446 	struct tdq *tdq;
1447 
1448 #ifdef SMP
1449 	sched_setup_smp();
1450 #else
1451 	tdq_setup(TDQ_SELF(), 0);
1452 #endif
1453 	tdq = TDQ_SELF();
1454 
1455 	/* Add thread0's load since it's running. */
1456 	TDQ_LOCK(tdq);
1457 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1458 	tdq_load_add(tdq, &thread0);
1459 	tdq->tdq_lowpri = thread0.td_priority;
1460 	TDQ_UNLOCK(tdq);
1461 }
1462 
1463 /*
1464  * This routine determines time constants after stathz and hz are setup.
1465  */
1466 /* ARGSUSED */
1467 static void
1468 sched_initticks(void *dummy)
1469 {
1470 	int incr;
1471 
1472 	realstathz = stathz ? stathz : hz;
1473 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1474 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1475 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1476 	    realstathz);
1477 
1478 	/*
1479 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1480 	 * hz not being evenly divisible by stathz on all platforms.
1481 	 */
1482 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1483 	/*
1484 	 * This does not work for values of stathz that are more than
1485 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1486 	 */
1487 	if (incr == 0)
1488 		incr = 1;
1489 	tickincr = incr;
1490 #ifdef SMP
1491 	/*
1492 	 * Set the default balance interval now that we know
1493 	 * what realstathz is.
1494 	 */
1495 	balance_interval = realstathz;
1496 	balance_ticks = balance_interval;
1497 	affinity = SCHED_AFFINITY_DEFAULT;
1498 #endif
1499 	if (sched_idlespinthresh < 0)
1500 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1501 }
1502 
1503 
1504 /*
1505  * This is the core of the interactivity algorithm.  Determines a score based
1506  * on past behavior.  It is the ratio of sleep time to run time scaled to
1507  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1508  * differs from the cpu usage because it does not account for time spent
1509  * waiting on a run-queue.  Would be prettier if we had floating point.
1510  *
1511  * When a thread's sleep time is greater than its run time the
1512  * calculation is:
1513  *
1514  *                           scaling factor
1515  * interactivity score =  ---------------------
1516  *                        sleep time / run time
1517  *
1518  *
1519  * When a thread's run time is greater than its sleep time the
1520  * calculation is:
1521  *
1522  *                           scaling factor
1523  * interactivity score =  ---------------------    + scaling factor
1524  *                        run time / sleep time
1525  */
1526 static int
1527 sched_interact_score(struct thread *td)
1528 {
1529 	struct td_sched *ts;
1530 	int div;
1531 
1532 	ts = td_get_sched(td);
1533 	/*
1534 	 * The score is only needed if this is likely to be an interactive
1535 	 * task.  Don't go through the expense of computing it if there's
1536 	 * no chance.
1537 	 */
1538 	if (sched_interact <= SCHED_INTERACT_HALF &&
1539 		ts->ts_runtime >= ts->ts_slptime)
1540 			return (SCHED_INTERACT_HALF);
1541 
1542 	if (ts->ts_runtime > ts->ts_slptime) {
1543 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1544 		return (SCHED_INTERACT_HALF +
1545 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1546 	}
1547 	if (ts->ts_slptime > ts->ts_runtime) {
1548 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1549 		return (ts->ts_runtime / div);
1550 	}
1551 	/* runtime == slptime */
1552 	if (ts->ts_runtime)
1553 		return (SCHED_INTERACT_HALF);
1554 
1555 	/*
1556 	 * This can happen if slptime and runtime are 0.
1557 	 */
1558 	return (0);
1559 
1560 }
1561 
1562 /*
1563  * Scale the scheduling priority according to the "interactivity" of this
1564  * process.
1565  */
1566 static void
1567 sched_priority(struct thread *td)
1568 {
1569 	int score;
1570 	int pri;
1571 
1572 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1573 		return;
1574 	/*
1575 	 * If the score is interactive we place the thread in the realtime
1576 	 * queue with a priority that is less than kernel and interrupt
1577 	 * priorities.  These threads are not subject to nice restrictions.
1578 	 *
1579 	 * Scores greater than this are placed on the normal timeshare queue
1580 	 * where the priority is partially decided by the most recent cpu
1581 	 * utilization and the rest is decided by nice value.
1582 	 *
1583 	 * The nice value of the process has a linear effect on the calculated
1584 	 * score.  Negative nice values make it easier for a thread to be
1585 	 * considered interactive.
1586 	 */
1587 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1588 	if (score < sched_interact) {
1589 		pri = PRI_MIN_INTERACT;
1590 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1591 		    sched_interact) * score;
1592 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1593 		    ("sched_priority: invalid interactive priority %d score %d",
1594 		    pri, score));
1595 	} else {
1596 		pri = SCHED_PRI_MIN;
1597 		if (td_get_sched(td)->ts_ticks)
1598 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1599 			    SCHED_PRI_RANGE - 1);
1600 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1601 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1602 		    ("sched_priority: invalid priority %d: nice %d, "
1603 		    "ticks %d ftick %d ltick %d tick pri %d",
1604 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1605 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1606 		    SCHED_PRI_TICKS(td_get_sched(td))));
1607 	}
1608 	sched_user_prio(td, pri);
1609 
1610 	return;
1611 }
1612 
1613 /*
1614  * This routine enforces a maximum limit on the amount of scheduling history
1615  * kept.  It is called after either the slptime or runtime is adjusted.  This
1616  * function is ugly due to integer math.
1617  */
1618 static void
1619 sched_interact_update(struct thread *td)
1620 {
1621 	struct td_sched *ts;
1622 	u_int sum;
1623 
1624 	ts = td_get_sched(td);
1625 	sum = ts->ts_runtime + ts->ts_slptime;
1626 	if (sum < SCHED_SLP_RUN_MAX)
1627 		return;
1628 	/*
1629 	 * This only happens from two places:
1630 	 * 1) We have added an unusual amount of run time from fork_exit.
1631 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1632 	 */
1633 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1634 		if (ts->ts_runtime > ts->ts_slptime) {
1635 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1636 			ts->ts_slptime = 1;
1637 		} else {
1638 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1639 			ts->ts_runtime = 1;
1640 		}
1641 		return;
1642 	}
1643 	/*
1644 	 * If we have exceeded by more than 1/5th then the algorithm below
1645 	 * will not bring us back into range.  Dividing by two here forces
1646 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1647 	 */
1648 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1649 		ts->ts_runtime /= 2;
1650 		ts->ts_slptime /= 2;
1651 		return;
1652 	}
1653 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1654 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1655 }
1656 
1657 /*
1658  * Scale back the interactivity history when a child thread is created.  The
1659  * history is inherited from the parent but the thread may behave totally
1660  * differently.  For example, a shell spawning a compiler process.  We want
1661  * to learn that the compiler is behaving badly very quickly.
1662  */
1663 static void
1664 sched_interact_fork(struct thread *td)
1665 {
1666 	struct td_sched *ts;
1667 	int ratio;
1668 	int sum;
1669 
1670 	ts = td_get_sched(td);
1671 	sum = ts->ts_runtime + ts->ts_slptime;
1672 	if (sum > SCHED_SLP_RUN_FORK) {
1673 		ratio = sum / SCHED_SLP_RUN_FORK;
1674 		ts->ts_runtime /= ratio;
1675 		ts->ts_slptime /= ratio;
1676 	}
1677 }
1678 
1679 /*
1680  * Called from proc0_init() to setup the scheduler fields.
1681  */
1682 void
1683 schedinit(void)
1684 {
1685 	struct td_sched *ts0;
1686 
1687 	/*
1688 	 * Set up the scheduler specific parts of thread0.
1689 	 */
1690 	ts0 = td_get_sched(&thread0);
1691 	ts0->ts_ltick = ticks;
1692 	ts0->ts_ftick = ticks;
1693 	ts0->ts_slice = 0;
1694 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1695 }
1696 
1697 /*
1698  * This is only somewhat accurate since given many processes of the same
1699  * priority they will switch when their slices run out, which will be
1700  * at most sched_slice stathz ticks.
1701  */
1702 int
1703 sched_rr_interval(void)
1704 {
1705 
1706 	/* Convert sched_slice from stathz to hz. */
1707 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1708 }
1709 
1710 /*
1711  * Update the percent cpu tracking information when it is requested or
1712  * the total history exceeds the maximum.  We keep a sliding history of
1713  * tick counts that slowly decays.  This is less precise than the 4BSD
1714  * mechanism since it happens with less regular and frequent events.
1715  */
1716 static void
1717 sched_pctcpu_update(struct td_sched *ts, int run)
1718 {
1719 	int t = ticks;
1720 
1721 	/*
1722 	 * The signed difference may be negative if the thread hasn't run for
1723 	 * over half of the ticks rollover period.
1724 	 */
1725 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1726 		ts->ts_ticks = 0;
1727 		ts->ts_ftick = t - SCHED_TICK_TARG;
1728 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1729 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1730 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1731 		ts->ts_ftick = t - SCHED_TICK_TARG;
1732 	}
1733 	if (run)
1734 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1735 	ts->ts_ltick = t;
1736 }
1737 
1738 /*
1739  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1740  * if necessary.  This is the back-end for several priority related
1741  * functions.
1742  */
1743 static void
1744 sched_thread_priority(struct thread *td, u_char prio)
1745 {
1746 	struct td_sched *ts;
1747 	struct tdq *tdq;
1748 	int oldpri;
1749 
1750 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1751 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1752 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1753 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1754 	if (td != curthread && prio < td->td_priority) {
1755 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1756 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1757 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1758 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1759 		    curthread);
1760 	}
1761 	ts = td_get_sched(td);
1762 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1763 	if (td->td_priority == prio)
1764 		return;
1765 	/*
1766 	 * If the priority has been elevated due to priority
1767 	 * propagation, we may have to move ourselves to a new
1768 	 * queue.  This could be optimized to not re-add in some
1769 	 * cases.
1770 	 */
1771 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1772 		sched_rem(td);
1773 		td->td_priority = prio;
1774 		sched_add(td, SRQ_BORROWING);
1775 		return;
1776 	}
1777 	/*
1778 	 * If the thread is currently running we may have to adjust the lowpri
1779 	 * information so other cpus are aware of our current priority.
1780 	 */
1781 	if (TD_IS_RUNNING(td)) {
1782 		tdq = TDQ_CPU(ts->ts_cpu);
1783 		oldpri = td->td_priority;
1784 		td->td_priority = prio;
1785 		if (prio < tdq->tdq_lowpri)
1786 			tdq->tdq_lowpri = prio;
1787 		else if (tdq->tdq_lowpri == oldpri)
1788 			tdq_setlowpri(tdq, td);
1789 		return;
1790 	}
1791 	td->td_priority = prio;
1792 }
1793 
1794 /*
1795  * Update a thread's priority when it is lent another thread's
1796  * priority.
1797  */
1798 void
1799 sched_lend_prio(struct thread *td, u_char prio)
1800 {
1801 
1802 	td->td_flags |= TDF_BORROWING;
1803 	sched_thread_priority(td, prio);
1804 }
1805 
1806 /*
1807  * Restore a thread's priority when priority propagation is
1808  * over.  The prio argument is the minimum priority the thread
1809  * needs to have to satisfy other possible priority lending
1810  * requests.  If the thread's regular priority is less
1811  * important than prio, the thread will keep a priority boost
1812  * of prio.
1813  */
1814 void
1815 sched_unlend_prio(struct thread *td, u_char prio)
1816 {
1817 	u_char base_pri;
1818 
1819 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1820 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1821 		base_pri = td->td_user_pri;
1822 	else
1823 		base_pri = td->td_base_pri;
1824 	if (prio >= base_pri) {
1825 		td->td_flags &= ~TDF_BORROWING;
1826 		sched_thread_priority(td, base_pri);
1827 	} else
1828 		sched_lend_prio(td, prio);
1829 }
1830 
1831 /*
1832  * Standard entry for setting the priority to an absolute value.
1833  */
1834 void
1835 sched_prio(struct thread *td, u_char prio)
1836 {
1837 	u_char oldprio;
1838 
1839 	/* First, update the base priority. */
1840 	td->td_base_pri = prio;
1841 
1842 	/*
1843 	 * If the thread is borrowing another thread's priority, don't
1844 	 * ever lower the priority.
1845 	 */
1846 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1847 		return;
1848 
1849 	/* Change the real priority. */
1850 	oldprio = td->td_priority;
1851 	sched_thread_priority(td, prio);
1852 
1853 	/*
1854 	 * If the thread is on a turnstile, then let the turnstile update
1855 	 * its state.
1856 	 */
1857 	if (TD_ON_LOCK(td) && oldprio != prio)
1858 		turnstile_adjust(td, oldprio);
1859 }
1860 
1861 /*
1862  * Set the base user priority, does not effect current running priority.
1863  */
1864 void
1865 sched_user_prio(struct thread *td, u_char prio)
1866 {
1867 
1868 	td->td_base_user_pri = prio;
1869 	if (td->td_lend_user_pri <= prio)
1870 		return;
1871 	td->td_user_pri = prio;
1872 }
1873 
1874 void
1875 sched_lend_user_prio(struct thread *td, u_char prio)
1876 {
1877 
1878 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1879 	td->td_lend_user_pri = prio;
1880 	td->td_user_pri = min(prio, td->td_base_user_pri);
1881 	if (td->td_priority > td->td_user_pri)
1882 		sched_prio(td, td->td_user_pri);
1883 	else if (td->td_priority != td->td_user_pri)
1884 		td->td_flags |= TDF_NEEDRESCHED;
1885 }
1886 
1887 /*
1888  * Like the above but first check if there is anything to do.
1889  */
1890 void
1891 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1892 {
1893 
1894 	if (td->td_lend_user_pri != prio)
1895 		goto lend;
1896 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
1897 		goto lend;
1898 	if (td->td_priority >= td->td_user_pri)
1899 		goto lend;
1900 	return;
1901 
1902 lend:
1903 	thread_lock(td);
1904 	sched_lend_user_prio(td, prio);
1905 	thread_unlock(td);
1906 }
1907 
1908 #ifdef SMP
1909 /*
1910  * This tdq is about to idle.  Try to steal a thread from another CPU before
1911  * choosing the idle thread.
1912  */
1913 static void
1914 tdq_trysteal(struct tdq *tdq)
1915 {
1916 	struct cpu_group *cg;
1917 	struct tdq *steal;
1918 	cpuset_t mask;
1919 	int cpu, i;
1920 
1921 	if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1922 		return;
1923 	CPU_FILL(&mask);
1924 	CPU_CLR(PCPU_GET(cpuid), &mask);
1925 	/* We don't want to be preempted while we're iterating. */
1926 	spinlock_enter();
1927 	TDQ_UNLOCK(tdq);
1928 	for (i = 1, cg = tdq->tdq_cg; ; ) {
1929 		cpu = sched_highest(cg, mask, steal_thresh);
1930 		/*
1931 		 * If a thread was added while interrupts were disabled don't
1932 		 * steal one here.
1933 		 */
1934 		if (tdq->tdq_load > 0) {
1935 			TDQ_LOCK(tdq);
1936 			break;
1937 		}
1938 		if (cpu == -1) {
1939 			i++;
1940 			cg = cg->cg_parent;
1941 			if (cg == NULL || i > trysteal_limit) {
1942 				TDQ_LOCK(tdq);
1943 				break;
1944 			}
1945 			continue;
1946 		}
1947 		steal = TDQ_CPU(cpu);
1948 		/*
1949 		 * The data returned by sched_highest() is stale and
1950                  * the chosen CPU no longer has an eligible thread.
1951 		 */
1952 		if (steal->tdq_load < steal_thresh ||
1953 		    steal->tdq_transferable == 0)
1954 			continue;
1955 		tdq_lock_pair(tdq, steal);
1956 		/*
1957 		 * If we get to this point, unconditonally exit the loop
1958 		 * to bound the time spent in the critcal section.
1959 		 *
1960 		 * If a thread was added while interrupts were disabled don't
1961 		 * steal one here.
1962 		 */
1963 		if (tdq->tdq_load > 0) {
1964 			TDQ_UNLOCK(steal);
1965 			break;
1966 		}
1967 		/*
1968 		 * The data returned by sched_highest() is stale and
1969                  * the chosen CPU no longer has an eligible thread.
1970 		 */
1971 		if (steal->tdq_load < steal_thresh ||
1972 		    steal->tdq_transferable == 0) {
1973 			TDQ_UNLOCK(steal);
1974 			break;
1975 		}
1976 		/*
1977 		 * If we fail to acquire one due to affinity restrictions,
1978 		 * bail out and let the idle thread to a more complete search
1979 		 * outside of a critical section.
1980 		 */
1981 		if (tdq_move(steal, tdq) == NULL) {
1982 			TDQ_UNLOCK(steal);
1983 			break;
1984 		}
1985 		TDQ_UNLOCK(steal);
1986 		break;
1987 	}
1988 	spinlock_exit();
1989 }
1990 #endif
1991 
1992 /*
1993  * Handle migration from sched_switch().  This happens only for
1994  * cpu binding.
1995  */
1996 static struct mtx *
1997 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1998 {
1999 	struct tdq *tdn;
2000 
2001 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2002 	    "thread %s queued on absent CPU %d.", td->td_name,
2003 	    td_get_sched(td)->ts_cpu));
2004 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2005 #ifdef SMP
2006 	tdq_load_rem(tdq, td);
2007 	/*
2008 	 * Do the lock dance required to avoid LOR.  We grab an extra
2009 	 * spinlock nesting to prevent preemption while we're
2010 	 * not holding either run-queue lock.
2011 	 */
2012 	spinlock_enter();
2013 	thread_lock_block(td);	/* This releases the lock on tdq. */
2014 
2015 	/*
2016 	 * Acquire both run-queue locks before placing the thread on the new
2017 	 * run-queue to avoid deadlocks created by placing a thread with a
2018 	 * blocked lock on the run-queue of a remote processor.  The deadlock
2019 	 * occurs when a third processor attempts to lock the two queues in
2020 	 * question while the target processor is spinning with its own
2021 	 * run-queue lock held while waiting for the blocked lock to clear.
2022 	 */
2023 	tdq_lock_pair(tdn, tdq);
2024 	tdq_add(tdn, td, flags);
2025 	tdq_notify(tdn, td);
2026 	TDQ_UNLOCK(tdn);
2027 	spinlock_exit();
2028 #endif
2029 	return (TDQ_LOCKPTR(tdn));
2030 }
2031 
2032 /*
2033  * Variadic version of thread_lock_unblock() that does not assume td_lock
2034  * is blocked.
2035  */
2036 static inline void
2037 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2038 {
2039 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2040 	    (uintptr_t)mtx);
2041 }
2042 
2043 /*
2044  * Switch threads.  This function has to handle threads coming in while
2045  * blocked for some reason, running, or idle.  It also must deal with
2046  * migrating a thread from one queue to another as running threads may
2047  * be assigned elsewhere via binding.
2048  */
2049 void
2050 sched_switch(struct thread *td, struct thread *newtd, int flags)
2051 {
2052 	struct tdq *tdq;
2053 	struct td_sched *ts;
2054 	struct mtx *mtx;
2055 	int srqflag;
2056 	int cpuid, preempted;
2057 
2058 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2059 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2060 
2061 	cpuid = PCPU_GET(cpuid);
2062 	tdq = TDQ_SELF();
2063 	ts = td_get_sched(td);
2064 	mtx = td->td_lock;
2065 	sched_pctcpu_update(ts, 1);
2066 	ts->ts_rltick = ticks;
2067 	td->td_lastcpu = td->td_oncpu;
2068 	td->td_oncpu = NOCPU;
2069 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2070 	    (flags & SW_PREEMPT) != 0;
2071 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2072 	td->td_owepreempt = 0;
2073 	if (!TD_IS_IDLETHREAD(td))
2074 		tdq->tdq_switchcnt++;
2075 	/*
2076 	 * The lock pointer in an idle thread should never change.  Reset it
2077 	 * to CAN_RUN as well.
2078 	 */
2079 	if (TD_IS_IDLETHREAD(td)) {
2080 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2081 		TD_SET_CAN_RUN(td);
2082 	} else if (TD_IS_RUNNING(td)) {
2083 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2084 		srqflag = preempted ?
2085 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2086 		    SRQ_OURSELF|SRQ_YIELDING;
2087 #ifdef SMP
2088 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2089 			ts->ts_cpu = sched_pickcpu(td, 0);
2090 #endif
2091 		if (ts->ts_cpu == cpuid)
2092 			tdq_runq_add(tdq, td, srqflag);
2093 		else {
2094 			KASSERT(THREAD_CAN_MIGRATE(td) ||
2095 			    (ts->ts_flags & TSF_BOUND) != 0,
2096 			    ("Thread %p shouldn't migrate", td));
2097 			mtx = sched_switch_migrate(tdq, td, srqflag);
2098 		}
2099 	} else {
2100 		/* This thread must be going to sleep. */
2101 		TDQ_LOCK(tdq);
2102 		mtx = thread_lock_block(td);
2103 		tdq_load_rem(tdq, td);
2104 #ifdef SMP
2105 		if (tdq->tdq_load == 0)
2106 			tdq_trysteal(tdq);
2107 #endif
2108 	}
2109 
2110 #if (KTR_COMPILE & KTR_SCHED) != 0
2111 	if (TD_IS_IDLETHREAD(td))
2112 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2113 		    "prio:%d", td->td_priority);
2114 	else
2115 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2116 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2117 		    "lockname:\"%s\"", td->td_lockname);
2118 #endif
2119 
2120 	/*
2121 	 * We enter here with the thread blocked and assigned to the
2122 	 * appropriate cpu run-queue or sleep-queue and with the current
2123 	 * thread-queue locked.
2124 	 */
2125 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2126 	newtd = choosethread();
2127 	/*
2128 	 * Call the MD code to switch contexts if necessary.
2129 	 */
2130 	if (td != newtd) {
2131 #ifdef	HWPMC_HOOKS
2132 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2133 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2134 #endif
2135 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2136 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2137 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2138 		sched_pctcpu_update(td_get_sched(newtd), 0);
2139 
2140 #ifdef KDTRACE_HOOKS
2141 		/*
2142 		 * If DTrace has set the active vtime enum to anything
2143 		 * other than INACTIVE (0), then it should have set the
2144 		 * function to call.
2145 		 */
2146 		if (dtrace_vtime_active)
2147 			(*dtrace_vtime_switch_func)(newtd);
2148 #endif
2149 
2150 		cpu_switch(td, newtd, mtx);
2151 		/*
2152 		 * We may return from cpu_switch on a different cpu.  However,
2153 		 * we always return with td_lock pointing to the current cpu's
2154 		 * run queue lock.
2155 		 */
2156 		cpuid = PCPU_GET(cpuid);
2157 		tdq = TDQ_SELF();
2158 		lock_profile_obtain_lock_success(
2159 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2160 
2161 		SDT_PROBE0(sched, , , on__cpu);
2162 #ifdef	HWPMC_HOOKS
2163 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2164 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2165 #endif
2166 	} else {
2167 		thread_unblock_switch(td, mtx);
2168 		SDT_PROBE0(sched, , , remain__cpu);
2169 	}
2170 
2171 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2172 	    "prio:%d", td->td_priority);
2173 
2174 	/*
2175 	 * Assert that all went well and return.
2176 	 */
2177 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2178 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2179 	td->td_oncpu = cpuid;
2180 }
2181 
2182 /*
2183  * Adjust thread priorities as a result of a nice request.
2184  */
2185 void
2186 sched_nice(struct proc *p, int nice)
2187 {
2188 	struct thread *td;
2189 
2190 	PROC_LOCK_ASSERT(p, MA_OWNED);
2191 
2192 	p->p_nice = nice;
2193 	FOREACH_THREAD_IN_PROC(p, td) {
2194 		thread_lock(td);
2195 		sched_priority(td);
2196 		sched_prio(td, td->td_base_user_pri);
2197 		thread_unlock(td);
2198 	}
2199 }
2200 
2201 /*
2202  * Record the sleep time for the interactivity scorer.
2203  */
2204 void
2205 sched_sleep(struct thread *td, int prio)
2206 {
2207 
2208 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2209 
2210 	td->td_slptick = ticks;
2211 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2212 		td->td_flags |= TDF_CANSWAP;
2213 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2214 		return;
2215 	if (static_boost == 1 && prio)
2216 		sched_prio(td, prio);
2217 	else if (static_boost && td->td_priority > static_boost)
2218 		sched_prio(td, static_boost);
2219 }
2220 
2221 /*
2222  * Schedule a thread to resume execution and record how long it voluntarily
2223  * slept.  We also update the pctcpu, interactivity, and priority.
2224  */
2225 void
2226 sched_wakeup(struct thread *td)
2227 {
2228 	struct td_sched *ts;
2229 	int slptick;
2230 
2231 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2232 	ts = td_get_sched(td);
2233 	td->td_flags &= ~TDF_CANSWAP;
2234 	/*
2235 	 * If we slept for more than a tick update our interactivity and
2236 	 * priority.
2237 	 */
2238 	slptick = td->td_slptick;
2239 	td->td_slptick = 0;
2240 	if (slptick && slptick != ticks) {
2241 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2242 		sched_interact_update(td);
2243 		sched_pctcpu_update(ts, 0);
2244 	}
2245 	/*
2246 	 * Reset the slice value since we slept and advanced the round-robin.
2247 	 */
2248 	ts->ts_slice = 0;
2249 	sched_add(td, SRQ_BORING);
2250 }
2251 
2252 /*
2253  * Penalize the parent for creating a new child and initialize the child's
2254  * priority.
2255  */
2256 void
2257 sched_fork(struct thread *td, struct thread *child)
2258 {
2259 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2260 	sched_pctcpu_update(td_get_sched(td), 1);
2261 	sched_fork_thread(td, child);
2262 	/*
2263 	 * Penalize the parent and child for forking.
2264 	 */
2265 	sched_interact_fork(child);
2266 	sched_priority(child);
2267 	td_get_sched(td)->ts_runtime += tickincr;
2268 	sched_interact_update(td);
2269 	sched_priority(td);
2270 }
2271 
2272 /*
2273  * Fork a new thread, may be within the same process.
2274  */
2275 void
2276 sched_fork_thread(struct thread *td, struct thread *child)
2277 {
2278 	struct td_sched *ts;
2279 	struct td_sched *ts2;
2280 	struct tdq *tdq;
2281 
2282 	tdq = TDQ_SELF();
2283 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2284 	/*
2285 	 * Initialize child.
2286 	 */
2287 	ts = td_get_sched(td);
2288 	ts2 = td_get_sched(child);
2289 	child->td_oncpu = NOCPU;
2290 	child->td_lastcpu = NOCPU;
2291 	child->td_lock = TDQ_LOCKPTR(tdq);
2292 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2293 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2294 	ts2->ts_cpu = ts->ts_cpu;
2295 	ts2->ts_flags = 0;
2296 	/*
2297 	 * Grab our parents cpu estimation information.
2298 	 */
2299 	ts2->ts_ticks = ts->ts_ticks;
2300 	ts2->ts_ltick = ts->ts_ltick;
2301 	ts2->ts_ftick = ts->ts_ftick;
2302 	/*
2303 	 * Do not inherit any borrowed priority from the parent.
2304 	 */
2305 	child->td_priority = child->td_base_pri;
2306 	/*
2307 	 * And update interactivity score.
2308 	 */
2309 	ts2->ts_slptime = ts->ts_slptime;
2310 	ts2->ts_runtime = ts->ts_runtime;
2311 	/* Attempt to quickly learn interactivity. */
2312 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2313 #ifdef KTR
2314 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2315 #endif
2316 }
2317 
2318 /*
2319  * Adjust the priority class of a thread.
2320  */
2321 void
2322 sched_class(struct thread *td, int class)
2323 {
2324 
2325 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2326 	if (td->td_pri_class == class)
2327 		return;
2328 	td->td_pri_class = class;
2329 }
2330 
2331 /*
2332  * Return some of the child's priority and interactivity to the parent.
2333  */
2334 void
2335 sched_exit(struct proc *p, struct thread *child)
2336 {
2337 	struct thread *td;
2338 
2339 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2340 	    "prio:%d", child->td_priority);
2341 	PROC_LOCK_ASSERT(p, MA_OWNED);
2342 	td = FIRST_THREAD_IN_PROC(p);
2343 	sched_exit_thread(td, child);
2344 }
2345 
2346 /*
2347  * Penalize another thread for the time spent on this one.  This helps to
2348  * worsen the priority and interactivity of processes which schedule batch
2349  * jobs such as make.  This has little effect on the make process itself but
2350  * causes new processes spawned by it to receive worse scores immediately.
2351  */
2352 void
2353 sched_exit_thread(struct thread *td, struct thread *child)
2354 {
2355 
2356 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2357 	    "prio:%d", child->td_priority);
2358 	/*
2359 	 * Give the child's runtime to the parent without returning the
2360 	 * sleep time as a penalty to the parent.  This causes shells that
2361 	 * launch expensive things to mark their children as expensive.
2362 	 */
2363 	thread_lock(td);
2364 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2365 	sched_interact_update(td);
2366 	sched_priority(td);
2367 	thread_unlock(td);
2368 }
2369 
2370 void
2371 sched_preempt(struct thread *td)
2372 {
2373 	struct tdq *tdq;
2374 
2375 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2376 
2377 	thread_lock(td);
2378 	tdq = TDQ_SELF();
2379 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2380 	tdq->tdq_ipipending = 0;
2381 	if (td->td_priority > tdq->tdq_lowpri) {
2382 		int flags;
2383 
2384 		flags = SW_INVOL | SW_PREEMPT;
2385 		if (td->td_critnest > 1)
2386 			td->td_owepreempt = 1;
2387 		else if (TD_IS_IDLETHREAD(td))
2388 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2389 		else
2390 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2391 	}
2392 	thread_unlock(td);
2393 }
2394 
2395 /*
2396  * Fix priorities on return to user-space.  Priorities may be elevated due
2397  * to static priorities in msleep() or similar.
2398  */
2399 void
2400 sched_userret_slowpath(struct thread *td)
2401 {
2402 
2403 	thread_lock(td);
2404 	td->td_priority = td->td_user_pri;
2405 	td->td_base_pri = td->td_user_pri;
2406 	tdq_setlowpri(TDQ_SELF(), td);
2407 	thread_unlock(td);
2408 }
2409 
2410 /*
2411  * Handle a stathz tick.  This is really only relevant for timeshare
2412  * threads.
2413  */
2414 void
2415 sched_clock(struct thread *td)
2416 {
2417 	struct tdq *tdq;
2418 	struct td_sched *ts;
2419 
2420 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2421 	tdq = TDQ_SELF();
2422 #ifdef SMP
2423 	/*
2424 	 * We run the long term load balancer infrequently on the first cpu.
2425 	 */
2426 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2427 		if (balance_ticks && --balance_ticks == 0)
2428 			sched_balance();
2429 	}
2430 #endif
2431 	/*
2432 	 * Save the old switch count so we have a record of the last ticks
2433 	 * activity.   Initialize the new switch count based on our load.
2434 	 * If there is some activity seed it to reflect that.
2435 	 */
2436 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2437 	tdq->tdq_switchcnt = tdq->tdq_load;
2438 	/*
2439 	 * Advance the insert index once for each tick to ensure that all
2440 	 * threads get a chance to run.
2441 	 */
2442 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2443 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2444 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2445 			tdq->tdq_ridx = tdq->tdq_idx;
2446 	}
2447 	ts = td_get_sched(td);
2448 	sched_pctcpu_update(ts, 1);
2449 	if (td->td_pri_class & PRI_FIFO_BIT)
2450 		return;
2451 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2452 		/*
2453 		 * We used a tick; charge it to the thread so
2454 		 * that we can compute our interactivity.
2455 		 */
2456 		td_get_sched(td)->ts_runtime += tickincr;
2457 		sched_interact_update(td);
2458 		sched_priority(td);
2459 	}
2460 
2461 	/*
2462 	 * Force a context switch if the current thread has used up a full
2463 	 * time slice (default is 100ms).
2464 	 */
2465 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2466 		ts->ts_slice = 0;
2467 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2468 	}
2469 }
2470 
2471 u_int
2472 sched_estcpu(struct thread *td __unused)
2473 {
2474 
2475 	return (0);
2476 }
2477 
2478 /*
2479  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2480  * cooperative idle threads.
2481  */
2482 int
2483 sched_runnable(void)
2484 {
2485 	struct tdq *tdq;
2486 	int load;
2487 
2488 	load = 1;
2489 
2490 	tdq = TDQ_SELF();
2491 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2492 		if (tdq->tdq_load > 0)
2493 			goto out;
2494 	} else
2495 		if (tdq->tdq_load - 1 > 0)
2496 			goto out;
2497 	load = 0;
2498 out:
2499 	return (load);
2500 }
2501 
2502 /*
2503  * Choose the highest priority thread to run.  The thread is removed from
2504  * the run-queue while running however the load remains.  For SMP we set
2505  * the tdq in the global idle bitmask if it idles here.
2506  */
2507 struct thread *
2508 sched_choose(void)
2509 {
2510 	struct thread *td;
2511 	struct tdq *tdq;
2512 
2513 	tdq = TDQ_SELF();
2514 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2515 	td = tdq_choose(tdq);
2516 	if (td) {
2517 		tdq_runq_rem(tdq, td);
2518 		tdq->tdq_lowpri = td->td_priority;
2519 		return (td);
2520 	}
2521 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2522 	return (PCPU_GET(idlethread));
2523 }
2524 
2525 /*
2526  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2527  * we always request it once we exit a critical section.
2528  */
2529 static inline void
2530 sched_setpreempt(struct thread *td)
2531 {
2532 	struct thread *ctd;
2533 	int cpri;
2534 	int pri;
2535 
2536 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2537 
2538 	ctd = curthread;
2539 	pri = td->td_priority;
2540 	cpri = ctd->td_priority;
2541 	if (pri < cpri)
2542 		ctd->td_flags |= TDF_NEEDRESCHED;
2543 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2544 		return;
2545 	if (!sched_shouldpreempt(pri, cpri, 0))
2546 		return;
2547 	ctd->td_owepreempt = 1;
2548 }
2549 
2550 /*
2551  * Add a thread to a thread queue.  Select the appropriate runq and add the
2552  * thread to it.  This is the internal function called when the tdq is
2553  * predetermined.
2554  */
2555 void
2556 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2557 {
2558 
2559 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2560 	KASSERT((td->td_inhibitors == 0),
2561 	    ("sched_add: trying to run inhibited thread"));
2562 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2563 	    ("sched_add: bad thread state"));
2564 	KASSERT(td->td_flags & TDF_INMEM,
2565 	    ("sched_add: thread swapped out"));
2566 
2567 	if (td->td_priority < tdq->tdq_lowpri)
2568 		tdq->tdq_lowpri = td->td_priority;
2569 	tdq_runq_add(tdq, td, flags);
2570 	tdq_load_add(tdq, td);
2571 }
2572 
2573 /*
2574  * Select the target thread queue and add a thread to it.  Request
2575  * preemption or IPI a remote processor if required.
2576  */
2577 void
2578 sched_add(struct thread *td, int flags)
2579 {
2580 	struct tdq *tdq;
2581 #ifdef SMP
2582 	int cpu;
2583 #endif
2584 
2585 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2586 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2587 	    sched_tdname(curthread));
2588 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2589 	    KTR_ATTR_LINKED, sched_tdname(td));
2590 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2591 	    flags & SRQ_PREEMPTED);
2592 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2593 	/*
2594 	 * Recalculate the priority before we select the target cpu or
2595 	 * run-queue.
2596 	 */
2597 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2598 		sched_priority(td);
2599 #ifdef SMP
2600 	/*
2601 	 * Pick the destination cpu and if it isn't ours transfer to the
2602 	 * target cpu.
2603 	 */
2604 	cpu = sched_pickcpu(td, flags);
2605 	tdq = sched_setcpu(td, cpu, flags);
2606 	tdq_add(tdq, td, flags);
2607 	if (cpu != PCPU_GET(cpuid)) {
2608 		tdq_notify(tdq, td);
2609 		return;
2610 	}
2611 #else
2612 	tdq = TDQ_SELF();
2613 	TDQ_LOCK(tdq);
2614 	/*
2615 	 * Now that the thread is moving to the run-queue, set the lock
2616 	 * to the scheduler's lock.
2617 	 */
2618 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2619 	tdq_add(tdq, td, flags);
2620 #endif
2621 	if (!(flags & SRQ_YIELDING))
2622 		sched_setpreempt(td);
2623 }
2624 
2625 /*
2626  * Remove a thread from a run-queue without running it.  This is used
2627  * when we're stealing a thread from a remote queue.  Otherwise all threads
2628  * exit by calling sched_exit_thread() and sched_throw() themselves.
2629  */
2630 void
2631 sched_rem(struct thread *td)
2632 {
2633 	struct tdq *tdq;
2634 
2635 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2636 	    "prio:%d", td->td_priority);
2637 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2638 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2639 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2640 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2641 	KASSERT(TD_ON_RUNQ(td),
2642 	    ("sched_rem: thread not on run queue"));
2643 	tdq_runq_rem(tdq, td);
2644 	tdq_load_rem(tdq, td);
2645 	TD_SET_CAN_RUN(td);
2646 	if (td->td_priority == tdq->tdq_lowpri)
2647 		tdq_setlowpri(tdq, NULL);
2648 }
2649 
2650 /*
2651  * Fetch cpu utilization information.  Updates on demand.
2652  */
2653 fixpt_t
2654 sched_pctcpu(struct thread *td)
2655 {
2656 	fixpt_t pctcpu;
2657 	struct td_sched *ts;
2658 
2659 	pctcpu = 0;
2660 	ts = td_get_sched(td);
2661 
2662 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2663 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2664 	if (ts->ts_ticks) {
2665 		int rtick;
2666 
2667 		/* How many rtick per second ? */
2668 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2669 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2670 	}
2671 
2672 	return (pctcpu);
2673 }
2674 
2675 /*
2676  * Enforce affinity settings for a thread.  Called after adjustments to
2677  * cpumask.
2678  */
2679 void
2680 sched_affinity(struct thread *td)
2681 {
2682 #ifdef SMP
2683 	struct td_sched *ts;
2684 
2685 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2686 	ts = td_get_sched(td);
2687 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2688 		return;
2689 	if (TD_ON_RUNQ(td)) {
2690 		sched_rem(td);
2691 		sched_add(td, SRQ_BORING);
2692 		return;
2693 	}
2694 	if (!TD_IS_RUNNING(td))
2695 		return;
2696 	/*
2697 	 * Force a switch before returning to userspace.  If the
2698 	 * target thread is not running locally send an ipi to force
2699 	 * the issue.
2700 	 */
2701 	td->td_flags |= TDF_NEEDRESCHED;
2702 	if (td != curthread)
2703 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2704 #endif
2705 }
2706 
2707 /*
2708  * Bind a thread to a target cpu.
2709  */
2710 void
2711 sched_bind(struct thread *td, int cpu)
2712 {
2713 	struct td_sched *ts;
2714 
2715 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2716 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2717 	ts = td_get_sched(td);
2718 	if (ts->ts_flags & TSF_BOUND)
2719 		sched_unbind(td);
2720 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2721 	ts->ts_flags |= TSF_BOUND;
2722 	sched_pin();
2723 	if (PCPU_GET(cpuid) == cpu)
2724 		return;
2725 	ts->ts_cpu = cpu;
2726 	/* When we return from mi_switch we'll be on the correct cpu. */
2727 	mi_switch(SW_VOL, NULL);
2728 }
2729 
2730 /*
2731  * Release a bound thread.
2732  */
2733 void
2734 sched_unbind(struct thread *td)
2735 {
2736 	struct td_sched *ts;
2737 
2738 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2739 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2740 	ts = td_get_sched(td);
2741 	if ((ts->ts_flags & TSF_BOUND) == 0)
2742 		return;
2743 	ts->ts_flags &= ~TSF_BOUND;
2744 	sched_unpin();
2745 }
2746 
2747 int
2748 sched_is_bound(struct thread *td)
2749 {
2750 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2751 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2752 }
2753 
2754 /*
2755  * Basic yield call.
2756  */
2757 void
2758 sched_relinquish(struct thread *td)
2759 {
2760 	thread_lock(td);
2761 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2762 	thread_unlock(td);
2763 }
2764 
2765 /*
2766  * Return the total system load.
2767  */
2768 int
2769 sched_load(void)
2770 {
2771 #ifdef SMP
2772 	int total;
2773 	int i;
2774 
2775 	total = 0;
2776 	CPU_FOREACH(i)
2777 		total += TDQ_CPU(i)->tdq_sysload;
2778 	return (total);
2779 #else
2780 	return (TDQ_SELF()->tdq_sysload);
2781 #endif
2782 }
2783 
2784 int
2785 sched_sizeof_proc(void)
2786 {
2787 	return (sizeof(struct proc));
2788 }
2789 
2790 int
2791 sched_sizeof_thread(void)
2792 {
2793 	return (sizeof(struct thread) + sizeof(struct td_sched));
2794 }
2795 
2796 #ifdef SMP
2797 #define	TDQ_IDLESPIN(tdq)						\
2798     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2799 #else
2800 #define	TDQ_IDLESPIN(tdq)	1
2801 #endif
2802 
2803 /*
2804  * The actual idle process.
2805  */
2806 void
2807 sched_idletd(void *dummy)
2808 {
2809 	struct thread *td;
2810 	struct tdq *tdq;
2811 	int oldswitchcnt, switchcnt;
2812 	int i;
2813 
2814 	mtx_assert(&Giant, MA_NOTOWNED);
2815 	td = curthread;
2816 	tdq = TDQ_SELF();
2817 	THREAD_NO_SLEEPING();
2818 	oldswitchcnt = -1;
2819 	for (;;) {
2820 		if (tdq->tdq_load) {
2821 			thread_lock(td);
2822 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2823 			thread_unlock(td);
2824 		}
2825 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2826 #ifdef SMP
2827 		if (always_steal || switchcnt != oldswitchcnt) {
2828 			oldswitchcnt = switchcnt;
2829 			if (tdq_idled(tdq) == 0)
2830 				continue;
2831 		}
2832 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2833 #else
2834 		oldswitchcnt = switchcnt;
2835 #endif
2836 		/*
2837 		 * If we're switching very frequently, spin while checking
2838 		 * for load rather than entering a low power state that
2839 		 * may require an IPI.  However, don't do any busy
2840 		 * loops while on SMT machines as this simply steals
2841 		 * cycles from cores doing useful work.
2842 		 */
2843 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2844 			for (i = 0; i < sched_idlespins; i++) {
2845 				if (tdq->tdq_load)
2846 					break;
2847 				cpu_spinwait();
2848 			}
2849 		}
2850 
2851 		/* If there was context switch during spin, restart it. */
2852 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2853 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2854 			continue;
2855 
2856 		/* Run main MD idle handler. */
2857 		tdq->tdq_cpu_idle = 1;
2858 		/*
2859 		 * Make sure that tdq_cpu_idle update is globally visible
2860 		 * before cpu_idle() read tdq_load.  The order is important
2861 		 * to avoid race with tdq_notify.
2862 		 */
2863 		atomic_thread_fence_seq_cst();
2864 		/*
2865 		 * Checking for again after the fence picks up assigned
2866 		 * threads often enough to make it worthwhile to do so in
2867 		 * order to avoid calling cpu_idle().
2868 		 */
2869 		if (tdq->tdq_load != 0) {
2870 			tdq->tdq_cpu_idle = 0;
2871 			continue;
2872 		}
2873 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2874 		tdq->tdq_cpu_idle = 0;
2875 
2876 		/*
2877 		 * Account thread-less hardware interrupts and
2878 		 * other wakeup reasons equal to context switches.
2879 		 */
2880 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2881 		if (switchcnt != oldswitchcnt)
2882 			continue;
2883 		tdq->tdq_switchcnt++;
2884 		oldswitchcnt++;
2885 	}
2886 }
2887 
2888 /*
2889  * A CPU is entering for the first time or a thread is exiting.
2890  */
2891 void
2892 sched_throw(struct thread *td)
2893 {
2894 	struct thread *newtd;
2895 	struct tdq *tdq;
2896 
2897 	if (td == NULL) {
2898 #ifdef SMP
2899 		PCPU_SET(sched, DPCPU_PTR(tdq));
2900 #endif
2901 		/* Correct spinlock nesting and acquire the correct lock. */
2902 		tdq = TDQ_SELF();
2903 		TDQ_LOCK(tdq);
2904 		spinlock_exit();
2905 		PCPU_SET(switchtime, cpu_ticks());
2906 		PCPU_SET(switchticks, ticks);
2907 	} else {
2908 		tdq = TDQ_SELF();
2909 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2910 		tdq_load_rem(tdq, td);
2911 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2912 		td->td_lastcpu = td->td_oncpu;
2913 		td->td_oncpu = NOCPU;
2914 	}
2915 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2916 	newtd = choosethread();
2917 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2918 	cpu_throw(td, newtd);		/* doesn't return */
2919 }
2920 
2921 /*
2922  * This is called from fork_exit().  Just acquire the correct locks and
2923  * let fork do the rest of the work.
2924  */
2925 void
2926 sched_fork_exit(struct thread *td)
2927 {
2928 	struct tdq *tdq;
2929 	int cpuid;
2930 
2931 	/*
2932 	 * Finish setting up thread glue so that it begins execution in a
2933 	 * non-nested critical section with the scheduler lock held.
2934 	 */
2935 	cpuid = PCPU_GET(cpuid);
2936 	tdq = TDQ_SELF();
2937 	if (TD_IS_IDLETHREAD(td))
2938 		td->td_lock = TDQ_LOCKPTR(tdq);
2939 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2940 	td->td_oncpu = cpuid;
2941 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2942 	lock_profile_obtain_lock_success(
2943 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2944 
2945 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2946 	    "prio:%d", td->td_priority);
2947 	SDT_PROBE0(sched, , , on__cpu);
2948 }
2949 
2950 /*
2951  * Create on first use to catch odd startup conditons.
2952  */
2953 char *
2954 sched_tdname(struct thread *td)
2955 {
2956 #ifdef KTR
2957 	struct td_sched *ts;
2958 
2959 	ts = td_get_sched(td);
2960 	if (ts->ts_name[0] == '\0')
2961 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2962 		    "%s tid %d", td->td_name, td->td_tid);
2963 	return (ts->ts_name);
2964 #else
2965 	return (td->td_name);
2966 #endif
2967 }
2968 
2969 #ifdef KTR
2970 void
2971 sched_clear_tdname(struct thread *td)
2972 {
2973 	struct td_sched *ts;
2974 
2975 	ts = td_get_sched(td);
2976 	ts->ts_name[0] = '\0';
2977 }
2978 #endif
2979 
2980 #ifdef SMP
2981 
2982 /*
2983  * Build the CPU topology dump string. Is recursively called to collect
2984  * the topology tree.
2985  */
2986 static int
2987 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2988     int indent)
2989 {
2990 	char cpusetbuf[CPUSETBUFSIZ];
2991 	int i, first;
2992 
2993 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2994 	    "", 1 + indent / 2, cg->cg_level);
2995 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2996 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2997 	first = TRUE;
2998 	for (i = 0; i < MAXCPU; i++) {
2999 		if (CPU_ISSET(i, &cg->cg_mask)) {
3000 			if (!first)
3001 				sbuf_printf(sb, ", ");
3002 			else
3003 				first = FALSE;
3004 			sbuf_printf(sb, "%d", i);
3005 		}
3006 	}
3007 	sbuf_printf(sb, "</cpu>\n");
3008 
3009 	if (cg->cg_flags != 0) {
3010 		sbuf_printf(sb, "%*s <flags>", indent, "");
3011 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3012 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3013 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3014 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3015 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3016 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3017 		sbuf_printf(sb, "</flags>\n");
3018 	}
3019 
3020 	if (cg->cg_children > 0) {
3021 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3022 		for (i = 0; i < cg->cg_children; i++)
3023 			sysctl_kern_sched_topology_spec_internal(sb,
3024 			    &cg->cg_child[i], indent+2);
3025 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3026 	}
3027 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3028 	return (0);
3029 }
3030 
3031 /*
3032  * Sysctl handler for retrieving topology dump. It's a wrapper for
3033  * the recursive sysctl_kern_smp_topology_spec_internal().
3034  */
3035 static int
3036 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3037 {
3038 	struct sbuf *topo;
3039 	int err;
3040 
3041 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3042 
3043 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3044 	if (topo == NULL)
3045 		return (ENOMEM);
3046 
3047 	sbuf_printf(topo, "<groups>\n");
3048 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3049 	sbuf_printf(topo, "</groups>\n");
3050 
3051 	if (err == 0) {
3052 		err = sbuf_finish(topo);
3053 	}
3054 	sbuf_delete(topo);
3055 	return (err);
3056 }
3057 
3058 #endif
3059 
3060 static int
3061 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3062 {
3063 	int error, new_val, period;
3064 
3065 	period = 1000000 / realstathz;
3066 	new_val = period * sched_slice;
3067 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3068 	if (error != 0 || req->newptr == NULL)
3069 		return (error);
3070 	if (new_val <= 0)
3071 		return (EINVAL);
3072 	sched_slice = imax(1, (new_val + period / 2) / period);
3073 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3074 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3075 	    realstathz);
3076 	return (0);
3077 }
3078 
3079 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3080 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3081     "Scheduler name");
3082 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3083     NULL, 0, sysctl_kern_quantum, "I",
3084     "Quantum for timeshare threads in microseconds");
3085 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3086     "Quantum for timeshare threads in stathz ticks");
3087 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3088     "Interactivity score threshold");
3089 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3090     &preempt_thresh, 0,
3091     "Maximal (lowest) priority for preemption");
3092 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3093     "Assign static kernel priorities to sleeping threads");
3094 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3095     "Number of times idle thread will spin waiting for new work");
3096 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3097     &sched_idlespinthresh, 0,
3098     "Threshold before we will permit idle thread spinning");
3099 #ifdef SMP
3100 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3101     "Number of hz ticks to keep thread affinity for");
3102 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3103     "Enables the long-term load balancer");
3104 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3105     &balance_interval, 0,
3106     "Average period in stathz ticks to run the long-term balancer");
3107 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3108     "Attempts to steal work from other cores before idling");
3109 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3110     "Minimum load on remote CPU before we'll steal");
3111 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3112     0, "Topological distance limit for stealing threads in sched_switch()");
3113 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3114     "Always run the stealer from the idle thread");
3115 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3116     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3117     "XML dump of detected CPU topology");
3118 #endif
3119 
3120 /* ps compat.  All cpu percentages from ULE are weighted. */
3121 static int ccpu = 0;
3122 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
3123